This public-safe synthetic dataset contains 800 governed agent trajectories
for supervised tool-use training, 120 development tasks, and a separately
frozen set of 150 final evaluation tasks. It demonstrates how recorded
successful executions and reviewed safety
refusals can become training data without publishing user traces.
train_trajectories.jsonl: 800 training-only conversational tool-use rows… See the full description on the dataset page:
https://huggingface.co/datasets/zwright/hermes-flight-recorder-self-improving-agent-trajectories.
